Machine vision-assisted method for detecting the repair status of optic nerve damage

Abstract

This invention discloses a machine vision-assisted method for detecting the repair status of optic nerve injury, belonging to the field of medical informatics technology. The method includes: determining a bimodal image through interaction with a front-end device group, triggering a state detector, performing multi-scale image reconstruction and A2 polarization inference under dual-path image feature extraction, determining an A2 polarization probability heatmap, assessing the repair status and providing directional detection guidance, performing multiple rounds of detection analysis based on first-order global detection and second-order directional detection within the repair cycle until the end of the repair cycle, and generating a periodic state evolution map. This invention addresses the technical problem in existing technologies where detection is one-sided and fails to effectively apply repair correlation information at different levels, resulting in limited detection intelligence and accuracy of repair status assessment. By adopting a systematic detection method that can integrate multimodal imaging information and has multi-scale analysis and intelligent reasoning capabilities, the method can effectively improve the intelligence and accuracy of detection.

Description

Technical Field

[0001] This invention relates to the field of medical informatics technology, specifically to a machine vision-assisted method for detecting the state of optic nerve injury repair. Background Technology

[0002] Optic nerve injury is a major cause of irreversible visual impairment, and its repair process involves multiple levels of changes, including metabolic regulation, glial cell response, neural circuit remodeling, and behavioral function recovery.

[0003] Current assessments of optic nerve injury repair primarily rely on methods such as tissue sections, immunofluorescence staining, single-modal imaging, or behavioral tests. These methods typically suffer from drawbacks such as high invasiveness, limited detection dimensions, and insufficient temporal continuity, making it difficult to dynamically, precisely, and quantitatively assess the repair status throughout the repair cycle. Some studies have attempted to introduce multimodal imaging or machine learning analysis, but these are mostly limited to simple feature fusion or static classification, and have not yet effectively characterized the intrinsic relationship between metabolic state, cell phenotype, and functional recovery.

[0004] In summary, existing technologies for detecting the repair status of optic nerve injury still suffer from limitations in detection scope and failure to effectively utilize repair correlation information at different levels, resulting in limitations in the intelligence of detection and the accuracy of repair status assessment. Summary of the Invention

[0005] This application provides a machine vision-assisted method for detecting the repair status of optic nerve injury, which addresses the technical problem in the prior art that the detection is one-sided and fails to effectively apply repair correlation information at different levels, resulting in limited intelligence of detection and accuracy of repair status assessment.

[0006] In view of the above problems, this application provides a machine vision-assisted method for detecting the repair status of optic nerve damage.

[0007] This application provides a machine vision-assisted method for detecting the repair state of optic nerve injury. The method includes: determining a bimodal image through interaction with a front-end device group, wherein the bimodal image includes a first metabolic image and a second phenotypic image; receiving the bimodal image through an intelligent detection platform, triggering an embedded state detector, performing multi-scale graph reconstruction and A2 polarization inference under dual-path image feature extraction, and determining an A2 polarization probability heatmap, wherein the multi-scale reconstruction abstracts multi-level biological entity features in the dual-path graph features into nodes, and performs cross-scale feature aggregation of nodes based on spatial and functional relationships; the A2 polarization inference is the inference of the transformation of astrocytes from the A1 phenotype to the A2 phenotype; evaluating the repair state and providing directional detection guidance based on the A2 polarization probability heatmap, performing multiple rounds of detection analysis based on first-order global detection and second-order directional detection within the repair cycle until the end of the repair cycle, and generating a periodic state evolution map.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0009] The machine vision-assisted optic nerve injury repair status detection method provided in this application, through interaction with a front-end device group, determines a bimodal image. The intelligent detection platform receives the bimodal image, triggers the embedded state detector, and performs multi-scale graph reconstruction and A2 polarization inference under dual-path image feature extraction to determine an A2 polarization probability heatmap. Based on the A2 polarization probability heatmap, the repair status is evaluated and directional detection guidance is provided. During the repair cycle, multiple rounds of detection analysis based on first-order global detection and second-order directional detection are performed until the end of the repair cycle, generating a periodic state evolution map. This method addresses the technical problem in existing technologies where detection is one-sided and fails to effectively apply repair correlation information at different levels, resulting in limited detection intelligence and accuracy of repair status assessment. By adopting a systematic detection method that can integrate multimodal imaging information and has multi-scale analysis and intelligent reasoning capabilities, the method can effectively improve the intelligence and accuracy of detection. Attached Figure Description

[0010] Figure 1 A schematic diagram of the machine vision-assisted optic nerve injury repair status detection method provided in this application.

[0011] Figure 2 A schematic diagram of the process for obtaining the A2 polarization probability heatmap in the machine vision-assisted optic nerve injury repair state detection method provided in this application. Detailed Implementation

[0012] This application provides a machine vision-assisted method for detecting the repair status of optic nerve damage, which addresses the technical problems in existing technologies where detection is one-sided and fails to effectively utilize repair correlation information at different levels, resulting in limited intelligence in detection and accuracy in assessing repair status.

[0013] Example: Figure 1 As shown, this application provides a machine vision-assisted method for detecting the repair status of optic nerve damage, the method comprising:

[0014] S1: By interacting with the front-end device group, a bimodal image is determined, wherein the bimodal image includes a first metabolic image and a second phenotypic image.

[0015] The first metabolic image is a label-free optical metabolic imaging mode, and the first metabolic image is a high-dimensional image reflecting mitochondrial function and ketone body metabolism level; the second phenotypic image is a multi-channel fluorescence imaging mode in the same field of view, and the fluorescently labeled targets in the second phenotypic image include at least astrocytes, A2 polarization specific markers, and pixel regions of neuronal synapses and blood vessels.

[0016] In the implementation of this application, bimodal image data is acquired through data port communication between the intelligent detection platform and the front-end device group, ensuring the real-time performance and accuracy of the image data. The bimodal image consists of data from two imaging modes: a first metabolic image and a second phenotypic image, providing multi-dimensional information on the optic nerve injury repair status.

[0017] Specifically, the first metabolic image employs label-free optical metabolic imaging technology as its imaging mode, which can obtain metabolic information within the organism through optical signals such as reflection and scattering. In a feasible implementation of this embodiment, the acquisition of the first metabolic image can be achieved using label-free optical metabolic imaging equipment, such as a two-photon fluorescence lifetime microscope.

[0018] The first metabolic image reflects mitochondrial function and ketone body metabolism levels within the cell. For example, by specifically detecting the fluorescence lifetime of metabolic coenzymes such as NAD(P)H and FAD, or the chemical bond vibration signal of ketone body β-hydroxybutyrate, a high-dimensional image reflecting mitochondrial function and ketone body metabolism levels is generated, serving as the first metabolic image. Obtaining detailed metabolic information at the cellular level through the aforementioned imaging methods, especially changes in mitochondrial activity, is of great significance for assessing the repair status after optic nerve injury.

[0019] Simultaneously, the second phenotypic image is captured using multi-channel fluorescence imaging technology and is taken within the same field of view as the first metabolic image. In a feasible embodiment of this application, the acquisition of the second phenotypic image can be achieved using a multi-channel fluorescence microscopy device.

[0020] This invention employs multi-channel fluorescence imaging, utilizing the properties of fluorescent dyes excited by lasers of specific wavelengths to capture images of different intracellular structures or molecular markers. In this invention, the fluorescently labeled targets in the second phenotypic image include biological structures and functional features such as astrocytes, A2 polarization-specific markers, neuronal synapses, and blood vessels.

[0021] Astrocytes are a crucial cell type in the repair of nerve injuries, especially under A2 polarization, where their repair function is significantly enhanced. Therefore, fluorescent labeling of astrocytes can provide important information for assessing the repair status. A2 polarization-specific markers are molecular markers used to distinguish A2-polarized astrocytes. These markers can identify specific phenotypes of glial cells during the repair process, thereby assessing the repair status. Neuronal synapses and blood vessels are important structures in the nerve repair process. The remodeling and functional recovery of neuronal synapses, as well as angiogenesis and reconstitution, are crucial for optic nerve repair.

[0022] By acquiring and integrating the aforementioned bimodal images, this invention can comprehensively and accurately capture the dynamic changes of cells during the optic nerve repair process from both metabolic and phenotypic perspectives, providing a reliable data foundation for subsequent repair status assessment.

[0023] S2: The intelligent detection platform receives the dual-modal image, triggers the embedded state detector, performs multi-scale graph reconstruction and A2 polarization inference under dual-path image feature extraction, and determines the A2 polarization probability heatmap. The multi-scale reconstruction abstracts the multi-level biological entity features in the dual-path graph features into nodes, and performs cross-scale feature aggregation of nodes based on spatial and functional relationships. The A2 polarization inference is the inference of the transformation of astrocytes from the A1 phenotype to the A2 phenotype.

[0024] In this embodiment, the function of the state detector is to perform in-depth analysis based on the information in the dual-modal image and generate key indicators of the repair state. The process is as follows:

[0025] First, after receiving the dual-modal image data transmitted from the front-end device, the intelligent detection platform immediately triggers the embedded state detector.

[0026] In the state detector's workflow, specifically, the first metabolic image and the second phenotypic image are processed in parallel by being input into two branches of the dual-path feature extraction network, effectively capturing key information from different modal images and thus providing multi-dimensional data support for subsequent analysis.

[0027] Subsequently, the multi-scale map reconstruction stage begins. The goal of this stage is to reconstruct the first and second feature matrices obtained from the dual-path feature extraction network using multi-scale processing methods, generating a multi-scale map network. This allows for a more comprehensive reflection of the detailed changes in the image. The reconstructed multi-scale map network will contain information from multiple levels, including cellular, metabolic, and behavioral aspects, laying the foundation for subsequent inference and analysis.

[0028] Finally, the reconstructed multi-scale graph network is imported into the A2 polarization inference layer. Based on the data from the multi-scale graph network, after nonlinear processing by a hierarchical computational architecture, an A2 polarization probability heatmap is finally output. This heatmap reflects the A2 polarization state of different regions during the restoration process, and its variations can accurately indicate key regions and the state at different stages of optic nerve restoration.

[0029] Specifically, astrocytes, the most numerous glial cells in the central nervous system, are activated into reactive astrocytes after nerve injury. These cells can polarize into two phenotypes: A1 and A2. The A1 phenotype is neurotoxic, releasing inflammatory factors such as IL-1β and TNF-α, which exacerbate nerve damage. The A2 phenotype has a neuroprotective effect, secreting anti-inflammatory factors such as IL-10 and IL-13, as well as neurotrophic factors, which promote tissue repair.

[0030] In this application, the state of A2 polarization, that is, the transformation of astrocytes from the A1 phenotype to the A2 phenotype, is used to detect and evaluate the repair of visual impairment.

[0031] In summary, by deeply analyzing bimodal images and effectively combining metabolic information and phenotypic features, accurate assessment results of the repair status can be generated, which greatly improves the detection efficiency and accuracy in the optic nerve injury repair process.

[0032] Furthermore, before triggering the embedded state detector, the construction of the state detector, step S2 of this application includes:

[0033] A state detector is developed within the intelligent detection platform, and a communication connection is established between the data terminal of the intelligent detection platform and the front-end device group; wherein, the state detector includes a dual-path feature extraction network, a multi-scale reconstruction layer and an A2 polarization inference layer.

[0034] Furthermore, the A2 polarized inference layer is a layered computing architecture;

[0035] In this hierarchical computing architecture, the bottom layer represents the A2 polarization state, the middle layer represents the neuronal directional selection characteristics, and the top layer represents behavioral performance. The A2 polarization state is determined based on the conversion probability from the A1 phenotype to the A2 phenotype. The directional selection characteristics are the properties of reorganizing synaptic connections and restoring functional neural circuits during the repair process. The behavioral performance is the transformation of changes in metabolism and neural activity into the overt manifestation of functional recovery. Using the multi-scale graph network as the data source, the nonlinear process from the bottom to the top layer in the hierarchical computing architecture as the training target, and the A2 polarization probability distribution as the output target, the A2 polarization inference layer is generated through data-driven training. The nonlinear process represents the repair trajectory under the interaction relationship between metabolic state, neural activity, behavioral function, and the repair process.

[0036] In the implementation of this invention, the state detector is a multi-level and multi-functional component, whose core task is to automatically detect and evaluate the optic nerve injury repair status based on bimodal images.

[0037] To achieve this goal, a state detector first needs to be developed and deployed within the intelligent detection platform. Simultaneously, a communication connection needs to be established between the platform's data terminal and the front-end device group to ensure smooth transmission and real-time processing of image data. The front-end device group transmits dual-modal image data to the intelligent detection platform in real time, and the platform then performs subsequent intelligent inference and analysis tasks based on the data source.

[0038] The state detector comprises three main modules: a dual-path feature extraction network, a multi-scale reconstruction layer, and an A2 polarization inference layer. In this application, parallel feature extraction branches are used to form the architecture of the dual-path feature extraction network, which is then connected to the multi-scale reconstruction layer and the A2 polarization inference layer to determine the overall architecture of the state detector.

[0039] The operational logic of each layer of the architecture is as follows: the dual-path feature extraction network is responsible for extracting feature information of different modalities from the first metabolic image and the second phenotypic image, respectively, forming a first feature matrix and a second feature matrix. The multi-scale reconstruction layer further processes these feature matrices, reconstructing them into a multi-scale graph network through multi-dimensional information integration and data representation. This ensures that all features of the multimodal images are fully utilized, reflecting the repair status of the optic nerve from multiple aspects such as metabolism, cell phenotype, and behavior.

[0040] Specifically, the A2 polarization inference layer is a key component of the state detector. It employs a hierarchical computational architecture with multiple processing levels, each corresponding to different repair states from the bottom to the top. The bottom layer primarily processes A2 polarization states, a core marker in neural repair, reflecting the polarization changes of glial cells during the repair process. The middle layer focuses on the directional selection characteristics of neurons, i.e., how neurons reorganize their synaptic connections and restore functional neural circuits during repair. The top layer focuses on behavioral performance; its goal is to translate changes in metabolic and neural activity into explicit manifestations of functional recovery, i.e., observed behavioral features. This hierarchical architecture facilitates detailed analysis of the repair process at multiple levels, thus providing comprehensive information for repair state assessment.

[0041] In summary, based on the overall architecture and specific operating logic described above, a sample-driven training method is adopted to train until convergence, such as meeting the preset output accuracy, in order to obtain the constructed state detector.

[0042] Specifically, sample data is retrieved, that is, based on the above-mentioned bimodal type of image, the samples are identified according to the elements of each layer in the above logic, and converted into representations that conform to the above architecture logic sequence, which are used as the final training sample data to supervise the training of the above architecture.

[0043] Preferably, during the nonlinear computation of the A2 polarization inference layer, optimization is achieved through training based on the complex relationship between mapping metabolic state, neural activity, and behavioral function. Through data-driven training, the interaction between metabolic state and repair process is adaptively learned, and the A2 polarization probability distribution is output.

[0044] In summary, the state detector of the intelligent detection platform, by integrating modules such as dual-path feature extraction, multi-scale reconstruction, and A2 polarization inference, can achieve in-depth analysis and accurate inference of multi-dimensional data in the optic nerve repair process, generate high-precision repair state assessment results, and thus provide scientific support for the optimization of the repair process.

[0045] Furthermore, such as Figure 2 As shown, multi-scale map reconstruction and A2 polarization inference under dual-path image feature extraction are performed to determine the A2 polarization probability heatmap. Step S2 of this application includes:

[0046] The first metabolic image is input into the first branch of the dual-path feature extraction network, and the second phenotypic image is input into the second branch of the dual-path feature extraction network. Graph feature vectors are extracted in parallel to determine the first feature matrix and the second feature matrix. Based on the multi-scale reconstruction layer, the first feature matrix and the second feature matrix are reconstructed to determine the multi-scale graph network. The multi-scale graph network is then imported into the A2 polarization inference layer to output the A2 polarization probability heatmap.

[0047] In the implementation of this invention, the dual-path feature extraction network is responsible for extracting deep-level features from bimodal images. Specifically, the first metabolic image and the second phenotypic image are respectively input into the two branches of the dual-path feature extraction network for parallel processing. Each branch independently processes its corresponding image modality, thereby enabling the simultaneous extraction of feature information at different levels.

[0048] First, the first metabolic image is input into the first branch of the dual-path feature extraction network, which mainly extracts metabolic features that reflect mitochondrial function and ketone body metabolism levels, identifies key patterns in the metabolic state, and generates corresponding metabolic feature vectors, such as redox ratio and active ketone body metabolism regions, and integrates and outputs the first feature matrix.

[0049] Simultaneously, the second phenotypic image is input into the second branch of the dual-path feature extraction network. This branch focuses on processing the phenotypic image, extracting morphological and molecular features, specifically cellular and tissue features related to astrocytes, A2 polarization markers, neuronal synapses, and blood vessels. During this process, fluorescently labeled targets in the image are converted into image feature vectors, reflecting cellular behavior and structural changes during the repair process, and integrated to output the second feature matrix.

[0050] Next, the multi-scale reconstruction layer is processed, primarily involving the fusion and reconstruction of the first and second feature matrices. Specifically, this is achieved by integrating information from different scales to generate a multi-scale graphical network. In essence, the multi-scale reconstruction layer not only considers the performance of each feature matrix at a single scale but also combines metabolic and phenotypic features across scales, reflecting information from different scales during optic nerve repair from multiple dimensions. This gives the multi-scale graphical network higher expressive power, enabling it to more accurately capture the complex changes during the repair process.

[0051] Finally, the generated multi-scale graph network is input into the A2 polarization inference layer. The A2 polarization inference layer is responsible for in-depth inference and analysis of the input multi-scale graph network, and outputs an A2 polarization probability heatmap, which can show the changes in the A2 polarization state of different regions during the repair process.

[0052] By analyzing heatmaps, the dynamic status, key areas, and stages of the repair process can be clearly seen, thus providing a scientific basis for evaluating the repair effect.

[0053] In summary, by combining dual-path feature extraction, multi-scale reconstruction, and A2 polarization inference, this invention can deeply explore the metabolic state and phenotypic changes during the optic nerve injury repair process based on multimodal images, providing high-precision image analysis and inference support for real-time monitoring and evaluation of the repair process.

[0054] Furthermore, to determine the multi-scale graph network, step S2 of this application includes:

[0055] A molecular-metabolic scale map is constructed, wherein the peak concentration of metabolites is used as a node and spatial proximity and temporal correlation are used as edges; a cellular scale map is constructed, wherein individual astrocytes and neurons are used as nodes, morphological features are used as node attributes, and spatial contact and functional connectivity are used as edges; a tissue-behavioral scale map is constructed, wherein different time segments of a behavioral task are used as nodes; the molecular-metabolic scale map, cellular scale map and tissue-behavioral scale map are superimposed based on spatial phase to form a multi-scale graph network.

[0056] In the implementation of this invention, the multi-scale graph network integrates information from the molecular-metabolic scale, the cellular scale, and the tissue-behavioral scale. Through graph structures at different levels, it comprehensively reflects the dynamic changes during the optic nerve repair process. The following are the specific construction methods and key features of each scale graph.

[0057] First, a molecular-metabolic scale map is constructed. Peak metabolite concentrations are the core feature of each node, revealing metabolic activity during the repair process, particularly changes in mitochondrial function and ketone body metabolism. Using peak metabolite concentrations as nodes allows for precise localization of changes in different metabolites. Nodes are connected through spatial proximity and temporal correlation; that is, edges connect metabolites with spatial proximity and temporal correlation. Spatial proximity refers to the possibility that metabolites located close to each other in space may act together during repair, while temporal correlation indicates the relationship between metabolites with similar metabolic behaviors at adjacent time points. Edge connections based on spatial and temporal correlations help accurately capture the trajectory of metabolic activity changes, thereby revealing the metabolic dynamics during optic nerve injury repair.

[0058] Secondly, a cell-scale map is constructed. In this map, nodes represent individual astrocytes or neurons, which are important cell types during the repair process, especially during A2 polarization, where changes in astrocytes are crucial to the repair status. Each cell node not only represents the cell itself but also includes its morphological characteristics as node attributes. For example, morphological changes in astrocytes are closely related to their function; these morphological characteristics, such as the number of branches and volume, reflect different functional states of the cell during repair. Furthermore, the connections between cells are represented by edges, which represent spatial contact and functional connectivity, such as calcium signal synchronization. Spatial contact refers to the information transmission between cells through close proximity during repair, while functional connectivity describes the interaction and collaborative work of cells within neural circuits. In this way, the cell-scale map can describe in detail the interrelationships and dynamic changes of various cell types during the repair process.

[0059] Next, a tissue-behavioral scale map is constructed. The nodes in the tissue-behavioral scale map represent different time segments of a behavioral task, corresponding to different behavioral manifestations during the repair process, such as the recovery of neural activity and improvement of motor function. The construction of the tissue-behavioral scale map can link changes at the metabolic and cellular levels with changes in overt behavior, thus providing a more comprehensive assessment of the repair state. In this map, each node represents a specific stage of a behavioral task, while the edges reflect the interrelationships between different time segments.

[0060] In the aforementioned process, feature vector data corresponding to the construction of graphs at different scales are first extracted from the first feature matrix and the second feature matrix, and then the graph structure connection transformation is performed.

[0061] Finally, by using spatial phase, information from different scale maps is aligned spatially, ensuring that information from each scale can be effectively combined within the same spatial framework. This allows the finalized multi-scale map network to not only comprehensively describe the repair process at the metabolic, cellular, and behavioral levels, but also to provide a dynamic perspective across scales, revealing the complex changes in the optic nerve repair process.

[0062] In summary, by constructing a multi-scale graph network, this invention can simultaneously consider data from multiple dimensions, comprehensively reflecting the state of optic nerve repair from the molecular, cellular, and behavioral levels. The construction of this network provides rich, multi-dimensional, normalized data support for subsequent A2 polarization inference, helping to improve the accuracy and reliability of repair state assessment.

[0063] S3: Based on the A2 polarization probability heatmap, assess the repair status and provide directional detection guidance. During the repair cycle, perform multiple rounds of detection analysis based on first-order global detection and second-order directional detection until the repair cycle ends, generating a periodic state evolution map.

[0064] The system employs a dual-stage detection mode to manage multiple rounds of detection within the repair cycle. This dual-stage detection mode consists of a first-stage detection based on the entire domain and a second-stage detection based on a specific orientation.

[0065] In the implementation of this invention, the analysis of the A2 polarization probability heatmap reflects the key areas and repair progress during the optic nerve repair process. The intelligent detection platform can achieve dynamic evaluation of the repair status and continuously monitor the repair process by combining a dual-level detection mode.

[0066] First, the intelligent detection platform assesses the repair status based on the A2 polarization probability heatmap. The A2 polarization probability heatmap represents the spatial distribution and polarization evolution probability of glial cells' A2 polarization state during the repair process, particularly the polarization state of glial cells during optic nerve repair. By analyzing the spatial distribution of the A2 polarization probability heatmap, the status of different regions during the repair cycle and the trend of repair evolution can be accurately assessed. This assessment provides a theoretical basis for subsequent targeted detection.

[0067] Next, targeted testing guidance will be provided based on the assessment results. During the repair cycle, the goal of targeted testing is to select representative or anomalous areas based on the repair hotspots displayed in the A2 polarization probability heatmap for further in-depth testing. Targeted testing can focus on the most critical or potentially problematic areas during the repair process, ensuring effective use of testing resources and avoiding over-testing of irrelevant areas.

[0068] Preferably, to achieve comprehensive monitoring throughout the repair cycle, this invention employs a two-stage detection mode, comprising two parts: a global first-stage detection and a directional second-stage detection. Specifically, the first-stage detection is a global detection method covering the entire repair area. Through global data collection and analysis, it ensures a preliminary assessment of the entire repair process. The goal of this stage is to obtain a global view of the repair status and identify large-scale changes and potential problems during the repair process.

[0069] Following the completion of the first-order detection, the second-order targeted detection phase begins. Second-order detection is targeted, focusing on specific regions, particularly those showing high repair activity or abnormal behavior in the A2 polarization probability heatmap. Targeted detection involves selectively analyzing key or problematic areas in the repair process, thereby uncovering potential issues that may arise during the repair process. Second-order detection can efficiently identify the repair dynamics of local areas, providing a concrete basis for optimizing the repair strategy.

[0070] The combination of the above two-stage detection modes forms the multi-round detection mechanism proposed in this application. That is, throughout the entire repair cycle, first-stage global detection and second-stage directional detection are used as a group to conduct multiple detection analyses within the cycle, ensuring that the repair process is continuously tracked and accurately managed.

[0071] Finally, at the end of the repair cycle, by integrating multiple rounds of detection records and time-series data, the changes in the repair status are visualized, displaying the evolution trajectory of the repair status within the repair cycle and generating a periodic state evolution map. This clearly shows the repair progress at each stage of the repair process and provides a comprehensive evaluation of the repair effect. This not only provides an intuitive basis for summarizing the repair cycle but also provides valuable reference data for future repair work.

[0072] In summary, by employing a two-stage detection mode and multi-round detection management, this invention enables accurate tracking and evaluation of the repair status within the repair cycle. Combined with the A2 polarization probability heatmap, the platform can dynamically adjust the detection strategy, effectively improving the detection accuracy and management efficiency of the repair status.

[0073] Furthermore, step S3 of this application includes:

[0074] Identify the A2 polarization probability heatmap and assess the repair status, wherein the assessment is conducted jointly using spatial and temporal dimensions; based on the repair status, perform early warning management based on the repair process.

[0075] In the implementation of this invention, through in-depth analysis of the A2 polarization probability heatmap, the intelligent detection platform can dynamically assess the repair status and perform effective early warning management based on the assessment results.

[0076] First, the A2 polarization probability heatmap provides crucial spatial and temporal information about the repair status. By identifying different regions in the A2 polarization probability heatmap, the polarization state of each region during the repair process can be clearly understood. Specifically, higher polarization probability values ​​in the heatmap usually represent critical regions in the repair process; these regions may be in a state of rapid repair or anomaly. Lower polarization probability values ​​may indicate a slower repair process or poor repair results in that region.

[0077] When assessing the restoration status, the spatial dimension focuses on the restoration status of different areas, identifying which areas are at critical stages of restoration or have potential problems. The temporal dimension assessment, on the other hand, analyzes the changing trends of the restoration status based on time-series data throughout the restoration cycle. By jointly assessing both spatial and temporal dimensions, a comprehensive understanding of the evolution of the restoration process can be achieved, identifying patterns of change in the restoration status, thereby providing accurate predictions and feedback for the restoration progress.

[0078] Subsequently, early warning management based on the assessment results is implemented. The goal of early warning management is to identify potential problems in the remediation process as early as possible and take appropriate intervention measures.

[0079] Specifically, by monitoring the A2 polarization probability heatmap in real time, abnormal areas or stages in the repair process can be identified in a timely manner, and an early warning mechanism can be activated based on this information. For example, the early warning mechanism can set different thresholds according to different repair states. If the repair state is lower than the preset standard, the platform will automatically trigger an alarm and propose corresponding suggestions or action plans.

[0080] For example, in certain critical areas, if the A2 polarization state is low and the repair process is slow, an early warning can be issued to remind relevant personnel to conduct more detailed testing or intervention.

[0081] In summary, by identifying A2 polarization probability heatmaps and combining spatial and temporal dimensions for joint evaluation, the status of optic nerve repair can be comprehensively and dynamically assessed, enabling effective early warning management, timely detection of potential problems, and corresponding measures to ensure the smooth progress of the repair process.

[0082] Furthermore, step S3 of this application includes:

[0083] Based on the repair status, a directional detection guide is generated; based on the directional detection guide, the intelligent detection platform sends a directional detection instruction to the front-end device group to determine the directional detection image; the directional detection image is then sent back and preprocessed to the status detector to analyze the directional repair status.

[0084] In the implementation of this invention, generating directional detection guidance based on the repair status ensures that the repair process can be dynamically monitored efficiently and accurately, enabling targeted repair monitoring and optimizing the use of detection resources.

[0085] First, based on the assessment results of the repair status, a targeted inspection guide is generated. This guide, based on the analysis of the repair status, identifies which areas or stages require special attention during the current repair process. These might be areas with potential problems in the repair process, or areas where the repair effect is most significant and changes most rapidly. The targeted inspection guide thus represents the target areas that require focused monitoring. It allows for focusing on the most critical and challenging parts of the repair process, improving the targeting and efficiency of the inspection.

[0086] Next, based on the generated directional detection guidance, the intelligent detection platform sends directional detection instructions to the front-end device group. That is, the platform transmits information about key areas or stages requiring monitoring during the repair process to the front-end devices, instructing them to collect relevant directional detection images. By sending directional detection instructions, the front-end device group can focus on specific areas for image acquisition according to the platform's requirements, ensuring that the acquired data is highly consistent with the assessment of the repair status. The issuance of directional detection instructions further ensures the accuracy and real-time nature of repair status monitoring.

[0087] Subsequently, after acquiring the orientation detection images, the front-end equipment group transmits the images back to the intelligent detection platform. The transmitted images undergo preliminary preprocessing to ensure that the image quality meets the analysis requirements. The preprocessing process includes steps such as image denoising, contrast enhancement, image distortion correction, and effective portion segmentation based on orientation guidance to improve image usability and analysis accuracy. The preprocessed orientation detection images are then transmitted to the state detector for in-depth analysis. The specific analysis is based on the state detector's analysis process, with the same steps as above, to determine the orientation repair status.

[0088] In summary, the intelligent inspection platform enables detailed and targeted monitoring of the repair process, promptly identifying key areas and changes during the repair process. The generation and distribution of targeted inspection guidelines ensures the efficiency and relevance of repair status inspection, while avoiding over-inspection of non-critical areas and optimizing the use of inspection resources. This closed-loop mechanism not only improves the accuracy of repair status assessment but also ensures real-time tracking and dynamic adjustment of the repair process.

[0089] Furthermore, until the repair cycle ends, a periodic state evolution map is generated. Step S3 of this application includes:

[0090] For the repair cycle, multiple rounds of detection records are retrieved and integrated into a time series to determine the time-series detection records; the time-series detection records are converted into a periodic state evolution map and visualized on the display port of the intelligent detection platform.

[0091] In the implementation of this invention, retrieving and integrating multiple rounds of detection records and time series data for the repair cycle is a key step in the systematic and visual evaluation of the repair status. This provides continuous and dynamic feedback for the repair process and ensures long-term monitoring of the repair status.

[0092] First, regarding the repair cycle, the intelligent detection platform retrieves multiple rounds of detection records. These records, obtained through first-order global detection and second-order directional detection within the repair cycle, contain detailed data on the repair status at different time points. Each detection record includes information such as the assessment results of the repair area, changes in A2 polarization state, and extraction of cellular and metabolic characteristics. These records represent the state at different time points during the repair process, providing cross-timeframe data support for the repair progress.

[0093] Subsequently, time-series integration is performed, that is, the detection records at different time points are sorted and summarized in chronological order to form a continuous repair status trajectory, obtain a complete picture of the repair process, and ensure that the repair dynamics and progress of each area can be clearly understood at each stage of the repair cycle.

[0094] Next, the time-series detection records are converted into a periodic state evolution map. For example, the horizontal axis is used as the time axis, and each detection time node is used to determine the axial nodes, with the vertical axis representing the specific detection data. The periodic state evolution map is a dynamic display of the repair process, which can clearly reflect the repair progress at each time point within the repair cycle.

[0095] Preferably, the periodic state evolution map shows the repair status and key indicators, such as A2 polarization state, metabolic state, and cell behavior, in different regions and stages during the repair process.

[0096] Finally, on the intelligent detection platform's display port, users can intuitively and comprehensively view key changes during the repair process through a graphical and interactive interface. The display port can present the evolution of the repair status in the form of charts, heatmaps, timelines, etc., providing clear visual effects and helping users better understand the changes in the repair process.

[0097] The following is an exemplary implementation of a state detector based on an embodiment of this application:

[0098] Optionally, the state detector adopts a three-tiered heterogeneous network architecture. The input layer receives dual-modal images, where the first metabolic image input size is 512×512×3, and its three channels correspond to different metabolite signal intensities; the second phenotypic image input size is 512×512×4, and its four channels are labeled GFAP, S100A10, NeuN, and DAPI, respectively. The feature extraction stage employs two independent first and second branches based on a dual-path feature extraction network, both based on the ResNet-50 architecture. Preferably, a spatial attention layer is added after the third and fourth residual blocks to focus on directional features. The first branch outputs a 256-dimensional metabolic fingerprint feature vector, and the second branch outputs a 256-dimensional morphological feature vector, which serve as the first feature matrix and the second feature matrix in this application, respectively.

[0099] In the multi-scale graph reconstruction stage, based on the multi-scale reconstruction layer, individual cell instances are segmented from the second phenotypic image as basic nodes. Each node contains 64-dimensional features, composed of metabolic and morphological features of the corresponding region. During graph construction, the weights of the connections between nodes are based on three computable metrics: spatial Euclidean distance, with a threshold of 20 μm; metabolic activity correlation, with a Pearson correlation coefficient > 0.7; and morphological similarity, with a feature cosine similarity > 0.6. The number of graph nodes is pre-defined, specifically including the number of nodes in each scale of the molecular-metabolic scale graph, cell scale graph, and tissue-behavioral scale graph. Optionally, the LeakyReLU activation function is used. Attention analysis is performed on the first and second feature matrices at their respective scales. After construction, the spatial phases of the graphs at each scale are superimposed to form the constructed multi-scale graph network.

[0100] In the A2 polarization inference layer, the hierarchical inference head consists of three fully connected layers, for example, 256→128→64→2, which output the probability value of each node belonging to the A2 polarization state. This can be executed using standard deep learning frameworks.

[0101] During the specific training phase, the training dataset contains 300 pairs of bimodal image sequences from different optic nerve injury patterns at different repair stages, and continuous observations on days 1, 3, 7, 14, and 28 after injury.

[0102] The annotation process employs a semi-automatic workflow: the outlines of A2-polarized astrocytes are manually annotated, using S100A10 strong positivity and GFAP+ as criteria to generate binary mask labels. Spatial registration then maps the annotations to the corresponding regions of the metabolic image. Each sample ultimately includes: the original bimodal image, a cell instance segmentation map, node-level A2 classification labels (0 / 1), and a repair status score. The score can be a continuous value based on behavioral tests, ranging from 0 to 1.

[0103] The overall loss function is a multi-task weighted sum. The classification loss uses binary cross-entropy with class weights, and a positive sample weight of 3.5 is set to address the low proportion of A2 cells. The regression loss uses smoothed L1 loss to predict repair scores. The graph structure loss uses contrastive loss to encourage nodes with similar metabolic patterns to be closer together in the embedding space.

[0104] During training, an initial learning rate is set, and cosine annealing scheduling is used for batch training until convergence. Key hyperparameters include: 8 graph attention heads, a dropout rate of 0.3, and a feature fusion temperature parameter τ = 0.07.

[0105] In the specific execution process, the A2 probability prediction value of each node is obtained through forward propagation, which can optionally be set between 0 and 1. The obtained probability prediction value is mapped back to the original image space through bilinear interpolation: first, the probability value is assigned to the corresponding pixel position according to the centroid coordinates of the cell corresponding to the node; then, a Gaussian kernel, such as σ=2, is used for spatial smoothing; finally, color mapping is applied to convert the probability value into a heatmap display.

[0106] Meanwhile, preferably, the output is a quantifiable repair index, which is calculated as the weighted average of the A2 probability prediction values ​​of all nodes, with the weight being the metabolic activity intensity of the cell corresponding to that node.

[0107] Using a test set, the following performance metrics were verified by comparing the state detector output with the test samples: node-level A2 classification accuracy, AUC value, and Pearson correlation coefficient between repair index prediction and actual behavioral scores. In existing technologies, error sources include feature confusion caused by inaccurate cell segmentation boundaries and misjudgments due to weak metabolic signals in the very early stages of damage. This application effectively improves AUC compared to single-modal analysis and effectively improves detection accuracy compared to fully connected networks.

[0108] The machine vision-assisted method for detecting the repair status of optic nerve injury provided in this application has the following technical advantages:

[0109] 1. Acquiring dual-modal images—metabolic and phenotypic images—simultaneously captures mitochondrial function, ketone body metabolism levels, and phenotypic features such as astrocytes and neuronal synapses, achieving multi-dimensional, comprehensive detection of the repair state and overcoming the limitations of single-modal detection. The state detector, through dual-path feature extraction, multi-scale map reconstruction, and A2 polarization inference, fuses repair information from different levels to accurately output an A2 polarization probability heatmap, significantly improving the accuracy and refinement of repair state assessment.

[0110] 2. Employing a detection mode combining first-order global detection and second-order directional detection, and combining the repair status to generate directional detection commands, the system focuses on in-depth analysis of key areas. Simultaneously, it visually presents the repair timeline changes through a periodic state evolution map, enabling dynamic tracking and precise control. The system automates image acquisition, feature extraction, multi-scale inference, and result visualization, reducing manual intervention and improving detection efficiency.

[0111] Through the foregoing detailed description of the machine vision-assisted optic nerve injury repair status detection method, those skilled in the art can clearly understand the machine vision-assisted optic nerve injury repair status detection method in this embodiment. As for the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and relevant parts can be referred to the method section description.

[0112] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A machine vision-assisted method for detecting the repair status of optic nerve damage, characterized in that, The method includes: A bimodal image is determined through interaction with the front-end device group, wherein the bimodal image includes a first metabolic image and a second phenotypic image; The intelligent detection platform receives the dual-modal image, triggers the embedded state detector, performs multi-scale graph reconstruction and A2 polarization inference under dual-path image feature extraction, and determines the A2 polarization probability heatmap. The multi-scale reconstruction abstracts the multi-level biological entity features in the dual-path graph features into nodes and aggregates the cross-scale features of the nodes according to the spatial and functional relationship. The A2 polarization inference is the inference of the transformation of astrocytes from the A1 phenotype to the A2 phenotype. Based on the A2 polarization probability heatmap, the repair status is evaluated and directional detection guidance is provided. During the repair cycle, multiple rounds of detection analysis based on first-order global detection and second-order directional detection are performed until the repair cycle ends, generating a periodic state evolution map. This includes performing multi-scale map reconstruction and A2 polarization inference under dual-path image feature extraction to determine the A2 polarization probability heatmap, including: The first metabolic image is input into the first branch of the dual-path feature extraction network, and the second phenotypic image is input into the second branch of the dual-path feature extraction network. Graph feature vectors are extracted in parallel to determine the first feature matrix and the second feature matrix. Based on the multi-scale reconstruction layer, the first feature matrix and the second feature matrix are reconstructed to determine the multi-scale graph network; The multi-scale graph network is imported into the A2 polarization inference layer to output an A2 polarization probability heatmap.

2. The machine vision-assisted method for detecting the repair status of optic nerve damage as described in claim 1, characterized in that, The first metabolic image is a label-free optical metabolic imaging mode, and the first metabolic image reflects a high-dimensional image of mitochondrial function and ketone body metabolism level; The second phenotypic image is imaged using multichannel fluorescence imaging in the same field of view. The fluorescently labeled targets in the second phenotypic image include astrocytes, A2 polarization-specific markers, and pixel regions of neuronal synapses and blood vessels.

3. The machine vision-assisted method for detecting the repair status of optic nerve damage as described in claim 1, characterized in that, Before triggering the embedded state detector, the construction of the state detector includes: Develop a status detector within the intelligent detection platform and establish a communication connection between the data terminal of the intelligent detection platform and the front-end device group; The state detector includes a dual-path feature extraction network, a multi-scale reconstruction layer, and an A2 polarization inference layer.

4. The machine vision-assisted method for detecting the repair status of optic nerve damage as described in claim 1, characterized in that, Determine the multi-scale graphical network, including: A molecular-metabolic scale map is constructed, wherein the molecular-metabolic scale map uses metabolite concentration peaks as nodes and spatial proximity and temporal correlation as edges; Construct a cell-scale map, wherein the cell-scale map uses individual astrocytes and neurons as nodes, morphological features as node attributes, and spatial contact and functional connection as edges; Construct an organization-behavior scale map, wherein the organization-behavior scale map uses different time segments of a behavioral task as nodes; The molecular-metabolic scale map, cellular scale map, and tissue-behavioral scale map are superimposed based on spatial phase to form a multi-scale map network.

5. The machine vision-assisted method for detecting the repair status of optic nerve damage as described in claim 1, characterized in that, The A2 polarized inference layer is a layered computing architecture; The hierarchical computing architecture consists of an A2 polarization state at the bottom layer, neuronal directional selection characteristics at the middle layer, and behavioral manifestations at the top layer. The A2 polarization state is determined based on the conversion probability from the A1 phenotype to the A2 phenotype. The directional selection characteristics are the characteristics of reorganizing synaptic connections and restoring functional neural circuits during the repair process. The behavioral manifestations are the transformation of changes in metabolism and neural activity into the overt manifestations of functional recovery. Using the multi-scale graph network as the data source, the nonlinear process from the bottom to the top layer in the hierarchical computing architecture as the training objective, and the A2 polarization probability distribution as the output objective, the A2 polarization inference layer is generated through data-driven training. The nonlinear process represents the repair trajectory under the interaction between metabolic state, neural activity, behavioral function, and repair process.

6. The machine vision-assisted method for detecting the repair status of optic nerve damage as described in claim 1, characterized in that, Identify the A2 polarization probability heatmap and assess the repair status, wherein the assessment is conducted jointly using spatial and temporal dimensions. Based on the repair status, early warning management is implemented based on the repair process.

7. The machine vision-assisted method for detecting the repair status of optic nerve damage as described in claim 6, characterized in that, Based on the repair status, a directional detection guide is generated; Based on the directional detection guidance, the intelligent detection platform sends directional detection instructions to the front-end device group to determine the directional detection image; The directional detection image is sent back and preprocessed to the state detector to analyze the directional repair status.

8. The machine vision-assisted method for detecting the repair status of optic nerve damage as described in claim 7, characterized in that, A dual-stage detection mode is adopted to perform multiple rounds of detection management within the repair cycle. The dual-stage detection mode consists of a first-stage detection based on the entire domain and a second-stage detection based on orientation.

9. The machine vision-assisted method for detecting the repair status of optic nerve damage as described in claim 8, characterized in that, Until the end of the repair cycle, a periodic state evolution map is generated, including: For the repair cycle, multiple rounds of detection records are retrieved and integrated into a time series to determine the time-series detection records; The time-series detection records are converted into periodic state evolution maps and visualized on the display port of the intelligent detection platform.

Images